MAI-Transcribe-2評測：砍價72%，但語言表現不一 | MAI-Transcribe-2 Review: 72% Cheaper, Mixed Accuracy

By Kit 小克 | AI Tool Observer | 2026-09-05

🇹🇼 MAI-Transcribe-2評測：砍價72%，但語言表現不一

MAI-Transcribe-2 是 Microsoft AI 在 2026 年 9 月推出的最新語音辨識模型，主打「全世界最快、最準、最便宜」，每小時轉錄只要 0.1 美元，比前代 MAI-Transcribe-1 的 0.36 美元砍了 72%。價格確實吸引人，但拆開 60 種語言逐一檢視後，「全球最準」的說法沒那麼一致。

MAI-Transcribe-2 是什麼？

MAI-Transcribe-2 是 Microsoft 自家的語音轉文字模型，已上架 Microsoft Foundry 與 MAI Playground，開發者今天就能透過 API 呼叫，鎖定客服逐字稿、會議紀錄、字幕產製這類大量轉錄場景。

價格：每小時 0.1 美元，但只到今年底

官方舉例：一家一年處理 10 萬小時客服錄音的銀行或電信商，帳單會從 3.6 萬美元降到 1 萬美元。但要注意，這是「限時到 2026 年底」的早鳥價，正式牌價還沒公布——現在把長期成本模型建立在 0.1 美元上，年底可能要重算一次。

準確度：FLEURS 排名第一，但「平均值」藏了陷阱

在涵蓋 60 種語言的 FLEURS 基準測試中，MAI-Transcribe-2 拿下第一，平均字詞錯誤率（WER）是 5.2%，等於每 20 個字錯 1 個；中文（4.5%）、法文（2.8%）表現亮眼。但這個平均值其實比前代 1.5 版的 3.7% 還高，因為這次擴大到低資源語言，拉低了整體平均。個別語言上，Google 的 Gemini 3.1 Pro 在亞美尼亞語（6.1% 對 12.6%）、丹麥語小勝，ElevenLabs 的 Scribe v2 在南非荷蘭語領先。你的語言未必落在它最強的那一邊。

怎麼用、該注意什麼

需透過 Microsoft Foundry 的 Speech 資源呼叫，支援 WAV、MP3、FLAC，檔案上限 300MB
預設多語模式，支援自動語言辨識與對話中途切換語言
正式導入前，先用自己實際的語言、口音、噪音條件測一輪，不要只看官方平均分數

對做客服逐字稿、會議記錄、Podcast 字幕的人來說，MAI-Transcribe-2 的價格很有吸引力，中英文使用者受益最直接。但「全球最準」背後是一份長長的語言排名表，你的語言未必排前面。好不好用，試了才知道。

🇺🇸 MAI-Transcribe-2 Review: 72% Cheaper, Mixed Accuracy

MAI-Transcribe-2 is Microsoft AI's new speech recognition model, launched in early September 2026 with a headline claim: fastest, most accurate, and cheapest transcription in the world. It costs just $0.10 per audio hour, a 72% cut from the $0.36 launch price of MAI-Transcribe-1. That number is real — but it's also temporary, and "most accurate" turns out to depend heavily on which language you're transcribing.

What MAI-Transcribe-2 Actually Is

MAI-Transcribe-2 is Microsoft's in-house speech-to-text model, now live on Microsoft Foundry and the MAI Playground, callable via API today. The pitch is simple: drive transcription cost down while staying near the top of accuracy leaderboards, aimed squarely at high-volume use cases like call-center transcripts, meeting notes, and subtitle generation.

Pricing: $0.10/Hour, But Only Through 2026

Microsoft's early-bird price is $0.10 per audio hour. Their own example: a bank or telecom processing 100,000 hours of call-center audio a year sees its bill drop from $36,000 to $10,000. That's a real saving — but two caveats matter. This price is explicitly limited to "through the end of 2026," and Microsoft hasn't announced what it reverts to afterward. If you're building a long-term cost model around $0.10/hour today, expect to redo the math in a few months.

Accuracy: #1 on FLEURS, But the Average Hides a Catch

On the FLEURS benchmark across 60 languages, MAI-Transcribe-2 ranks first with an average word error rate (WER) of 5.2% — roughly 1 word wrong in 20. On Artificial Analysis's non-streaming tests, it drops to 2.0%. Chinese (4.5% WER) and French (2.8% WER) both look strong.

But honestly, that 5.2% average is actually higher than MAI-Transcribe-1.5's 3.7% — because this release expanded coverage into more low-resource languages, which drags the overall average down. Look at individual languages and the picture gets messier: Google's Gemini 3.1 Pro beats it on Armenian (6.1% vs. 12.6%) and edges it on Danish, while ElevenLabs' Scribe v2 wins on Afrikaans. If your use case sits in one of those languages, MAI-Transcribe-2 may not be the best pick.

How to Try It, and What to Watch For

Requires a Microsoft Foundry Speech resource; supports WAV, MP3, and FLAC files up to 300MB
Defaults to multi-lingual mode with automatic language ID and mid-conversation code-switching
Before rolling it into production, test it against your actual languages, accents, and background noise — don't rely on the headline average

For anyone producing call-center transcripts, meeting notes, or podcast subtitles, the price alone makes MAI-Transcribe-2 worth a look — especially if you work in Chinese or English. But behind the "most accurate in the world" marketing sits a long per-language leaderboard, and your language might not be near the top. 好不好用，試了才知道 — the only way to know is to try it yourself.

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士